Dispatch (Claude Code)

Launch non-blocking Claude Code headless tasks from slash command dispatch. Use when user requests async coding jobs and does not require slash-only Claude p...

Security Scan

The skill is internally consistent with its purpose (launching asynchronous Claude Code jobs); it runs bundled scripts, reads a limited allowlisted env file, and starts background runs — but you should be aware of some operational dependencies and local-file/callback behaviours before installing.

Purpose & Capability

The skill's name/description match what the code does (dispatch headless/interactive Claude Code jobs). However the package metadata declares no required binaries while the scripts assume several runtime tools (python3, jq, tmux, script(1), and the Claude CLI binary). Not listing these dependencies is a functional mismatch the operator should be aware of.

Instruction Scope

SKILL.md simply delegates to scripts/run_dispatch.sh. The scripts' actions (validate args, create result dirs, write metadata, spawn a background process running the bundled dispatch.sh and claude wrapper) are consistent with the stated behaviour. The scripts read a local env file and may scan the workspace for a dispatch-callback.json to auto-detect callback settings; they do not themselves download remote code.

Install Mechanism

No install spec is provided (instruction-only with bundled scripts). There are no remote downloads or package installs in the bundle, and all code is included in the skill. This is the lower-risk install model.

Credentials

The runtime safely parses only allowlisted KEY=VALUE lines from an optional dispatch.env.local (no 'source'), matching the security disclosure. The allowed keys list is reasonable for the skill. That said, the skill does not declare a primary credential and the process will inherit any environment variables already present in the agent runtime (e.g., OPENCLAW_SESSION_KEY / OPENCLAW_GATEWAY_TOKEN if set externally). Operators should confirm what environment will be present, since sensitive gateway/session tokens in the parent env could be used by the scripts or downstream tools.

Persistence & Privilege

The skill is not always-enabled and does not modify other skills or system-wide config. It launches background processes and writes logs/results under configurable paths, which is expected for a dispatch/worker launcher.

Assessment

This skill appears to do what it says, but review these before installing: - Ensure required binaries exist on the host: python3, jq, tmux (for interactive mode), script(1) (optional), and the Claude CLI binary referenced by CLAUDE_CODE_BIN. The skill metadata does not list these, so you must provision them. - Inspect or create the optional env file (${OPENCLAW_DISPATCH_ENV:-...}/dispatch.env.local). The loader only exports allowlisted keys (no 'source'), so secrets in that file will be limited to the allowed set; however any sensitive env vars already set in the agent process may still be inherited. Remove or rotate sensitive tokens if you don't want them available. - By default network callbacks are not enabled (ENABLE_CALLBACK=0). If you enable callbacks, verify the workspace dispatch-callback.json and any OPENCLAW_* session/gateway settings to ensure results are sent only to intended destinations. - Be aware the dispatcher runs Claude Code inside the project workdir; if a dispatched job executes tests or runs repository code, that code will run on your host — avoid dispatching untrusted prompts against sensitive repositories. - If you want to test safely, use DISPATCH_DRY_RUN=1 to inspect the constructed command and result directory without launching the background job.

Run {baseDir}/scripts/run_dispatch.sh with user args.

Contract

• Format: /dispatch <project> <task-name> <prompt...>
• Workdir mapping: ${REPOS_ROOT:-/home/miniade/repos}/<project>
• Agent Teams policy: on-demand (enabled only if prompt contains Agent Team signals)
• Safety: headless runs enforce timeout via DISPATCH_TIMEOUT_SEC (default 7200s)

Local config

• optional env file: ${OPENCLAW_DISPATCH_ENV:-<workspace>/skills/dispatch.env.local}
• supports OpenClaw skills.entries.dispatch.env injection
• script is self-contained (bundled dispatch.sh + claude_code_run.py)

Security disclosure

• Reads only allowlisted env keys from dispatch.env.local using key=value parsing (no source).
• Starts a background local process (nohup) and writes logs/results under configured paths.
• Network callback is disabled by default; enable only with ENABLE_CALLBACK=1 and explicit group settings.
• Does not download remote code at runtime.

Behavior

1. Validate args and return usage if incomplete.
2. Start task in background (non-blocking).
3. Return one-line launch summary with run-id and log path.
4. Do not run extra validation unless requested.